Rotate capped training samples across iterations

scripts/build_failure_driven_yolo_sampling.py

@@ -7,6 +7,7 @@ import argparse
 import hashlib
 import json
 import math
+import re
 from collections import Counter
 from pathlib import Path
 from typing import Any
@@ -49,6 +50,7 @@ def build_sampling(
     context_positive_repeat: int = 5,
     context_negative_repeat: int = 6,
     max_region_share: float = 0.65,
+    sampling_round: int = 0,
 ) -> tuple[list[str], dict[str, Any]]:
     if assessment.get("status") != "continue_training_loop":
         raise ValueError("Failure-driven sampling requires a failed assessment")
@@ -62,6 +64,8 @@ def build_sampling(
         raise ValueError("Repeat factors must be positive")
     if not 0 < max_region_share <= 1:
         raise ValueError("max_region_share must be in (0, 1]")
+    if sampling_round < 0:
+        raise ValueError("sampling_round must be non-negative")
 
     samples = {item["sample_slug"]: item for item in manifest["samples"]}
     gates = assessment["gates"]
@@ -172,7 +176,13 @@ def build_sampling(
         base_paths = base_paths_by_region[region]
         extra_limit = capped_counts[region] - len(base_paths)
         image_paths.extend(base_paths)
-        image_paths.extend(extra_paths_by_region[region][:extra_limit])
+        extras = extra_paths_by_region[region]
+        if extras and extra_limit < len(extras):
+            # Rotate the capped repeat window between loop rounds so persistent
+            # failures cannot yield the exact same training list indefinitely.
+            offset = sampling_round % len(extras)
+            extras = extras[offset:] + extras[:offset]
+        image_paths.extend(extras[:extra_limit])
     repeat_counts = Counter({region: capped_counts[region] for region in capped_counts})
     if not image_paths:
         raise ValueError("No train-only tiles selected")
@@ -195,6 +205,7 @@ def build_sampling(
         "context_positive_repeat": context_positive_repeat,
         "context_negative_repeat": context_negative_repeat,
         "max_region_share": max_region_share,
+        "sampling_round": sampling_round,
         "source_train_tile_count": sum(
             1
             for tile in summary["tiles"]
@@ -223,11 +234,16 @@ def main() -> int:
     parser.add_argument("--context-positive-repeat", type=int, default=5)
     parser.add_argument("--context-negative-repeat", type=int, default=6)
     parser.add_argument("--max-region-share", type=float, default=0.65)
+    parser.add_argument("--sampling-round", type=int)
     args = parser.parse_args()
 
     summary = json.loads(args.summary.read_text(encoding="utf-8"))
     manifest = json.loads(args.corpus_manifest.read_text(encoding="utf-8"))
     assessment = json.loads(args.assessment.read_text(encoding="utf-8"))
+    sampling_round = args.sampling_round
+    if sampling_round is None:
+        match = re.fullmatch(r"iteration-(\d+)", args.output_dir.parent.name)
+        sampling_round = int(match.group(1)) if match else 0
     paths, metadata = build_sampling(
         summary=summary,
         manifest=manifest,
@@ -238,6 +254,7 @@ def main() -> int:
         context_positive_repeat=args.context_positive_repeat,
         context_negative_repeat=args.context_negative_repeat,
         max_region_share=args.max_region_share,
+        sampling_round=sampling_round,
     )
     args.output_dir.mkdir(parents=True, exist_ok=True)
     train_list = args.output_dir / "train-failure-driven.txt"

scripts/run_belgium_building_training_loop.py

@@ -164,6 +164,7 @@ def failure_sampling_command(
     corpus_manifest: Path,
     assessment: Path,
     output_dir: Path,
+    sampling_round: int = 0,
 ) -> list[str]:
     return [
         sys.executable,
@@ -172,6 +173,7 @@ def failure_sampling_command(
         "--corpus-manifest", str(corpus_manifest),
         "--assessment", str(assessment),
         "--output-dir", str(output_dir),
+        "--sampling-round", str(sampling_round),
     ]
 
 
@@ -443,6 +445,7 @@ def main() -> int:
                 corpus_manifest=args.corpus_manifest,
                 assessment=assessment,
                 output_dir=sampling_dir,
+                sampling_round=index,
             ),
             iteration_dir / "failure-driven-sampling.log",
         )